Different addressing modes?
Ulrich Weigand
weigand@immd1.informatik.uni-erlangen.de
Wed May 21 22:11:00 GMT 2003
Fred Cook wrote:
>I am still trying to get addressing modes correctly in my port.
>Basically my architecture has the following addressing modes:
>
> 1) absolute
> 2) register indirect
> 3) register + offset
> 4) register + register
I'm assuming this means the following RTX represent valid
memory operands:
1) (mem (symbol_ref ...))
2) (mem (reg ...))
3) (mem (plus (reg ...) (const_int ...)))
4) (mem (plus (reg ...) (reg ...)))
>I want to specify in my define_insn for "call" that 1+2 are allowed
>and in my define_insn for "movsi" that 2, 3, and 4 are allowed for
>loads and 2 and 3 for stores.
"call" can be handled specially. It is fine to have a pattern
that matches
(call (mem (symbol_ref ...))
even if (mem (symbol_ref ...)) is not a valid memory operand.
I would suggest to do the following:
- Define GO_IF_LEGITIMATE_ADDRESS so that 2), 3), and 4) above
are accepted. This means that usual memory accesses (i.e.
those for loads) simply use the "m" constraint.
- Define an additional constraint letter "S" to denote
'memory reference without a second register', and define
EXTRA_CONSTRAINTS to only accept 2) and 3) above for
this letter.
As this definition fulfils the requirements "a subset of
all memory references including all those consisting of
solely a base register", you can (and must) mark the
constraint letter "S" as EXTRA_MEMORY_CONSTRAINT. This
will cause reload to fix up the address if required
(note that you *must* provide a load-address type pattern
that accepts the full range of memory addresses as source).
With this set up, you can define movsi like this:
(define_insn "movsi"
[(set (match_operand:SI 0 "nonimmediate_operand" "=r,S")
(match_operand:SI 1 "general_operand" "m,r"))]
""
"@
ld %1 -> %0
st %0 %1"
)
All this is actually similar to the s390 setup, so you
might want to check out our backend for comparison ...
Bye,
Ulrich
--
Dr. Ulrich Weigand
weigand@informatik.uni-erlangen.de
More information about the Gcc
mailing list